Combined effects of genotype and childhood adversity shape variability of DNA methylation across age
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Czamara et al. Translational Psychiatry (2021)11:88
https://doi.org/10.1038/s41398-020-01147-z Translational Psychiatry
ARTICLE Open Access
Combined effects of genotype and childhood
adversity shape variability of DNA methylation
across age
Darina Czamara 1, Elleke Tissink 2, Johanna Tuhkanen3, Jade Martins 1, Yvonne Awaloff4, Amanda J. Drake 5
,
Batbayar Khulan5, Aarno Palotie6, Sibylle M. Winter7, Charles B. Nemeroff8, W. Edward Craighead 9,
Boadie W. Dunlop 9, Helen S. Mayberg9,10, Becky Kinkead 9, Sanjay J. Mathew 11, Dan V. Iosifescu10,12,
Thomas C. Neylan13, Christine M. Heim 14, Jari Lahti 3,15, Johan G. Eriksson16,17,18,19, Katri Räikkönen3,
Kerry J. Ressler 20, Nadine Provençal21,22 and Elisabeth B. Binder 1,9
Abstract
Lasting effects of adversity, such as exposure to childhood adversity (CA) on disease risk, may be embedded via
epigenetic mechanisms but findings from human studies investigating the main effects of such exposure on
epigenetic measures, including DNA methylation (DNAm), are inconsistent. Studies in perinatal tissues indicate that
variability of DNAm at birth is best explained by the joint effects of genotype and prenatal environment. Here, we
extend these analyses to postnatal stressors. We investigated the contribution of CA, cis genotype (G), and their
additive (G + CA) and interactive (G × CA) effects to DNAm variability in blood or saliva from five independent cohorts
1234567890():,;
1234567890():,;
1234567890():,;
1234567890():,;
with a total sample size of 1074 ranging in age from childhood to late adulthood. Of these, 541 were exposed to CA,
which was assessed retrospectively using self-reports or verified through social services and registries. For the majority
of sites (over 50%) in the adult cohorts, variability in DNAm was best explained by G + CA or G × CA but almost never
by CA alone. Across ages and tissues, 1672 DNAm sites showed consistency of the best model in all five cohorts, with
G × CA interactions explaining most variance. The consistent G × CA sites mapped to genes enriched in brain-specific
transcripts and Gene Ontology terms related to development and synaptic function. Interaction of CA with genotypes
showed the strongest contribution to DNAm variability, with stable effects across cohorts in functionally relevant
genes. This underscores the importance of including genotype in studies investigating the impact of environmental
factors on epigenetic marks.
Introduction later in life1–4. Exposure to CA is not only associated with
Childhood adversity (CA), including child abuse and disease risk, but also with a number of lasting biological
neglect, is a major risk factors for the development of and physiological changes, including alterations in brain
stress-related psychiatric and other medical disorders structure, function, and connectivity5, stress response6,
and immune function7.
DNA methylation (DNAm) has been proposed as a
Correspondence: Darina Czamara (darina@psych.mpg.de) or
biological process by which early-life adversity may have
Elisabeth B. Binder (binder@psych.mpg.de)
1
Department of Translational Research in Psychiatry, Max Planck Institute of lasting effects on gene transcription providing a molecular
Psychiatry, Kraepelinstr. 2-10, 80804 Munich, Germany
2
mechanism for how early environment could influence
Department of Complex Trait Genetics, Center for Neurogenomics and
health outcomes later in life8,9. A number of studies have
Cognitive Research, Amsterdam Neuroscience, Vrije Universiteit Amsterdam,
1081 HV Amsterdam, The Netherlands investigated DNAm changes with exposure to CA in
Full list of author information is available at the end of the article
© The Author(s) 2021
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction
in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if
changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain
permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.Czamara et al. Translational Psychiatry (2021)11:88 Page 2 of 11 peripheral tissues, such as saliva or blood, either using social services. This enabled us to test for the stability of candidate gene approaches or genome-wide DNAm stu- G and CA effects with age as well as across different dies (EWAS). Overall, while there is some evidence for the types of assessment of exposure. association between CA and altered patterns of DNAm, results for individual DNAm targets remain inconsistent10. Methods The majority of autosomal CpGs (about 80%) are not Samples variable across tissues and individuals11,12, leaving only Five independent cohorts were included in our analysis: about 20% of CpG sites that may contribute to differences GRADY, PReDICT, U19, BerlinLCS, and HBCS. All in phenotypes and health13. These variable CpGs are of subjects (or their legal guardians) gave written informed specific interest as they are enriched for functionally consent and ethical approval was given by the Institu- relevant genomic regions, associated with effects on gene tional Review Board or Ethical Committee of each site expression12. In contrast to CA, genetic factors have been participating in every study. Register linkage has been shown to have replicable influences on DNAm variability. conducted with permission from the register authority The impact of genetic variation, especially of single- (HBCS: the Finnish National Archives). nucleotide polymorphisms (SNPs), on DNAm in different The GRADY cohort consisted of 309 participants who tissues, has been investigated in many studies and a large were recruited as part of the GRADY Trauma Project at number of methylation quantitative trait loci (SNPs sig- the Grady Memorial Hospital in Atlanta, Georgia18. All nificantly associated with DNAm status14) have been participants come from an urban population with low discovered which are relatively stable throughout the life socioeconomic status and are characterized by high pre- course15. valence and severity of trauma over lifetime19–21. Environmental factors and genetic factors may thus act The PReDICT cohort consisted of 363 treatment-naive in concert to influence DNAm, however only a few studies patients who met criteria for current major depressive have investigated the joint effects of environment and disorder. All participants were recruited at three Atlanta genotype on DNAm variability. In the context of the sites associated with the Emory University School of influence of prenatal environments on DNAm at birth, Medicine, Department of Psychiatry and Behavioral Teh et al.16 as well as our group17 reported that combined Sciences22. effects of genotype and prenatal environment explain The U19 cohort consisted of 78 nonmedicated women most of the variance in umbilical cord and cord blood recruited at four academic sites in the USA (Emory DNAm. In fact, environment alone was almost never the University, Icahn School of Medicine at Mount Sinai, strongest driver, rather additive or interactive effects of Baylor College of Medicine, University of California San genotype (G) and environment (E) explained DNAm Francisco/San Francisco Veterans Affairs Medical Cen- variability best in the majority of CpGs. This may be ter). All U19 participants were untreated and had to fulfill specific of prenatal environments, where there is less time criteria for post-traumatic stress disorder (PTSD) for at for exposure. least 3 months23. Here, we aimed to expand the analysis of combined G The BerlinLCS cohort consisted of 173 children, who and E effects to a postnatal stressor (CA). We examined were recruited via child care centers, child and youth if, similar to our results in neonates, combined effects social services, child psychiatric departments, or pedia- are also stronger drivers of DNAm variation later in life tricians. Children were followed for 2.5 years with and if the proportion of explained variance varies with extensive psychometric and biological assessments. In time to exposure, i.e., whether effects of CA measured in addition, DNA from saliva samples was collected at five childhood are qualitatively or quantitatively different time points over the course of the study (every 6 months). than when measured later in life. For this purpose, we Cases were victims of one or more of the following: systematically tested main effects of CA (E = CA) and physical abuse, physical neglect, and/or emotional mal- genotype located in a 1 MB window of the CpG (G) on treatment (MT) requiring intervention by social services. DNAm as well as their additive (G + CA) and multi- The Helsinki Birth Cohort Study (HBCS)24–26 consisted plicative effects (G × CA). For each tested CpG site, we of 77 men who were evacuated to Sweden or Denmark sought the model that explained most of the DNAm unaccompanied by their parents during the Second World variability. We explored this in five independent cohorts War according to the Finnish National Archives’ regis- with a total of 1074 individuals, of whom 541 were ter27. The controls were 74 men who were not evacuated, exposed to CA. The five cohorts ranged in age from and who were matched to cases for birth year and father’s early childhood (3–5 years of age) to elderly individuals occupational status. These men donated blood for DNA (mean age of 64 years) with both retrospective self- samples in a clinical study in 2001–2004. Information on reports of CA and verified exposures by registries or these five cohorts is summarized in Table 1.
Czamara et al. Translational Psychiatry (2021)11:88 Page 3 of 11
Table 1 Demographic overview of cohorts.
Cohort Na Mean age (SD) Sex (male) Ethnicity Assessment of CA CA Na (%) Tissue Methylation array
GRADY 309 42.08 (12.92) 25.56% African American Self-report 148 (47.90%) Whole blood 450K
PReDICT 363 39.83 (11.50) 39.67% Mixed Self-report 164 (45.18%) Whole blood 450K
U19 78 39.27 (12.10) 0.00% Mixed Self-report 66 (84.62%) Whole blood 450K
BerlinLCS 173 4.23 (0.79) 52.60% Caucasian Documented 86 (49.71%) Saliva EPIC
HBCS 151 63.5 (2.8) 100% Finnish Documented 77 (51.00%) Whole blood 450K
In GRADY, PReDICT, and U19 CA refers to moderate-to-severe ranges of CTQ scores for either sexual, physical, or emotional abuse (GRADY n = 77, PReDICT n = 79, U19
n = 14) or to moderate-to-severe scores in at least two abuse groups (GRADY n = 71, PReDICT n = 85, U19 n = 52), in BerlinLCS CA refers to maltreatment and in HBCS
to evacuation and separation from parents in World War II.
SD standard deviation, CA childhood adversity.
a
N = sample size.
DNAm data (2%), significant deviation from the Hardy–Weinberg
Methylation450K BeadChips in GRADY, PReDICT, U19, equilibrium (HWE, p < 10−5), or a low minor allele fre-
and HBCS and by the Infinium MethylationEPIC BeadChip quency (MAF < 5%) were excluded from further analyses.
for BerlinLCS (for this study we focused on the baseline Afterward, additional SNPs were imputed using IMPUTE
methylation levels). Beta values were normalized using v239, the 1000 Genomes phase III sample served as
functional normalization28,29. Batch effects were removed reference panel40. Imputed SNPs with a low information
using ComBat30 with the sva package31. Subsequently, all content metric ( 95%, call rate > 95% for individuals and
tion33. Saliva cell counts for the BerlinLCS cohort were markers (99% for markers with MAF < 5%), MAF > 1%,
computed according to Smith et al.34. Smoking scores in and HWE p > 10−06. Moreover, heterozygosity, sex check,
each cohort were calculated as described by Elliott and relatedness checks were performed and any dis-
et al.35,36. For the BerlinLCS cohort, we computed a pre- crepancies were removed. Imputed genotype probabilities
natal smoking exposure according to Richmond et al.37. were converted into best-guessed genotypes using a
threshold of 0.90. SNPs were pruned to a reduced subset
Genotype data of approximately independent SNPs using a repeated
DNA isolation and SNP genotyping sliding window (window size of 100 kb, 5 kb shift at the
In all cohorts, except BerlinLCS, DNA was isolated from end of each step) procedure with a pairwise SNP R2
blood samples (GRADY: using either the ArchivePure threshold of 0.238.
DNA Blood Kit (5 Prime, Gaithersburg, MD, USA) or E.Z.
N.A. Mag-Bind Blood DNA Kit (Omega Bio-tek, Nor- Environmental data
cross, GA, USA), U19: using the PerkinElmer Chemagic Self-reported childhood trauma: childhood trauma
360 extraction robot). In BerlinLCS, DNA was isolated questionnaire (CTQ)
from saliva samples. Genome-wide SNP genotyping was The CTQ is a psychometrically validated assessment of
performed using Illumina OmniQuad (GRADY), Huma- physical, sexual, and emotional child abuse and neglect,
nOmniExpress BeadChips (PReDICT and U19), Illumina using 28 self-report items41. Participants in GRADY,
GSA-24 v2.0 BeadChips (BerlinLCS), and Illumina 610k PReDICT, and U19 were classified into three groups
chips (HBCS, modified Illumina 610k chip by the Well- based on established cutoff scores for moderate-to-severe
come Trust Sanger Institute, Cambridge, UK). exposure levels for each type of childhood abuse (CA; ≥10
for physical abuse, ≥8 for sexual abuse, ≥13 for emotional
Quality control and imputation abuse)41. The first group of participants scored in the
Quality control was performed in PLINK38 indepen- none-to-mild range for sexual, emotional, and physical
dently in all cohorts. Samples with low genotyping rate abuse and was classified as negative for exposure, theCzamara et al. Translational Psychiatry (2021)11:88 Page 4 of 11
second group scored in the moderate-to-severe range for Explaining variability of VMPs
either sexual, physical, or emotional abuse, and the third To assess to what extent genotype, environment (CA),
group scored moderate-to-severe in at least two abuse genotype and environment, as well as genotype–environment
groups. interaction contributed to variation in VMPs, four different
linear regression models (1–4) were tested for each over-
Verified childhood trauma lapping VMP in each cohort to identify the model with the
For BerlinLCS, maltreated (physical abuse, physical largest adjusted R2.
neglect, or emotional MT) children were recruited via (1) Environment model (CA): VMP ~ covariates + CA.
child welfare offices, child and youth social services, child (2) Genotype model (G): VMP ~ covariates + Gi.
psychiatric departments, or pediatricians and corrobora- (3) Additive model (G + CA): VMP ~ covariates + Gi
tion/details of MT exposure was obtained by caretaker + CA.
report. Assessment and coding of maltreatment was based (4) Interaction model (G × CA): VMP ~ covariates +
on42 Of the 173 children, 86 were victims of MT. The 87 Gi + CA + Gi × CA.
children in the control group presented with no MT, and VMP represents the uncorrected beta value of the
no other significant stressors as assessed with the pre- identified variable CpG site described above. For models
school age psychiatric assessment43. (2–4), Gi is a SNP-genotype coded by the minor allele
count (0, 1, 2); all pruned SNPs in a cis window of ±1 MB
Childhood adversities in the HBCS around the VMP were tested sequentially and the SNP
In the HBCS sample that was used for this study, 77 that presented with the largest adjusted R2 was selected
individuals had been evacuated to Sweden or Denmark for further analysis. Covariates are the DNAm con-
unaccompanied by their parents during the Second World founders and cohort-specific covariates described in the
War according to the Finnish National Archives’ register. preceding section. In models (3) and (4) all possible SNP
This experience was used as CA in comparison to the and CA combinations were tested
other 74 individuals of the sample, who had not been The model with the largest adjusted R2 value, explaining
evacuated. the most variance across (1)–(4), was chosen as the best
model for that VMP.
Statistical analyses A work flow of the general procedure is depicted in
All statistical analyses were performed in each cohort Supplementary Fig. 1.
independently using R version 3.5.2.
Mapping VMPs to genomic regions
Identification and characterization of overlapping variable VMPs were mapped to their genomic location using the R-
methylated CpGs packages minfi28 and ChIPseeker44 and to their correspond-
In order to correct DNAm levels for known con- ing ChromHMM states based on histone ChiP-Seq peaks
founders, these were regressed out of the beta values from the Roadmap Epigenomics project derived for blood
using linear regression separately for each cohort. Cov- cells (http://egg2.wustl.edu/roadmap/data/byFileType/peaks/
ariates were defined as sex, age, blood cell counts33, or consolidated/broadPeak/).
saliva cell counts for BerlinLCS34, smoking scores, and Enrichment tests were performed using Fisher’s tests.
genotype principal components to account for population The significance levels were set using Bonferroni correc-
stratification (GRADY: the first two PCs, PReDICT, and tion according to the number of performed tests.
U19 and BerlinLCS: the first five PCs, HBCS: the first
three PCs). Using residuals from these models, the med- Gene-set enrichment analysis
ian absolute deviation (MAD) was estimated per CpG as a VMP sites were mapped to their closest genes using the
robust measure of DNAm variability within each cohort. matchGenes function in the R-package bumphunter45.
The MAD score is preferred for this purpose as it is not Gene-set enrichments were tested using FUMA’s GEN-
driven by outliers. We tested the 80th, 85th, 90th, and E2FUNC v1.3.546 setting the FDR adjusted p values for
95th percentile as cutoffs of the MAD score. The 80th enrichment to 0.05 and considering Gene Ontology (GO)
percentile cutoff resulted in 45,962 variably methylated terms as well as tissue-specific transcripts derived from
probes (VMPs) overlapping among GRADY, PReDICT, GTEx v647. A minimum number of ten genes had to
and U19. These VMPs were selected for initial analysis overlap with the specific gene set. We compared enrich-
and defined as overlapping VMPs. For a sensitivity ana- ments for stable across age G × CA CpGs (variably
lysis, we further regressed out cohort-specific covariates methylated CpGs with the best model being G × CA
(anxiety score in U19; depression score in GRADY, PRe- across all cohorts, n = 1400) and stable in adults G × CA
DICT, and U19; PTSD score in GRADY and U19). CpGs (variably methylated CpGs with the best model
Addition of these covariates did not influence the results. being G × CA in the three adult cohorts, but not in theCzamara et al. Translational Psychiatry (2021)11:88 Page 5 of 11
Fig. 1 Overlapping VMPs between GRADY, PReDICT, and U19. Venn diagram of overlapping VMPs for GRADY, PReDICT, and U19 at a MAD score
threshold at the 80th percentile (a). Distribution of the best models explaining variation in DNAm across the three adult cohorts. Percentage of
overlapping VMPs (n = 45,962) best explained by G, CA, G + CA, or G × CA in each cohort using the highest adjusted R2 (b). Consistent best models
across three adult cohorts. Percentage of overlapping VMPs (n = 45,962) best explained by the same model type (G, CA, G + CA, or G × CA) in at least
two cohorts (c).
other two cohorts, n = 670). We used the group of effects of cis genotype (G) and CA together. For each
inconsistent CpGs (n = 5652) that showed different best cohort, we compared the adjusted R2 of four regression
models across all three adult cohorts as control. For the models (CA, G, G + CA, G × CA) to find the model which
enrichment, we created ten random subsets of genes best explained DNAm variation in VMPs. The adjusted R2
mapping to these inconsistent CpGs and equal in size to is well suited to determine the most predictive model as it
the number of genes mapping to stable across age G × CA adjusts for the number of parameters in the model and
CpGs (n = 1123 genes). only increases if the inclusion of these parameters also
increases the model fit48. In all cohorts, the majority of
Results VMPs was best explained by additive or interactive effects
VMPs in adults with self-reported retrospective CA of G and CA (see Fig. 1b and Supplementary Fig. 2 for
We first assessed which of the four models (CA, G, G + details) with CA alone being the best model only in very
CA, and G × CA) explained most of the DNAm variability few VMPs.
in the three adult cohorts (see Table 1) that used the CTQ Over 80% of overlapping VMPs showed a consistent
for retrospective assessment of CA. CpGs with a MAD best model across at least two of the three cohorts (see
score larger than the 80th MAD percentile and over- Fig. 1c and Supplementary Fig. 3). As we based our results
lapping between GRADY, PReDICT, and U19 were on pruned SNPs and as all cohorts had a different ethnic
defined as overlapping VMPs (n = 45,962 VMPs, see Fig. background, we matched the consistency based on best
1a). As previously described17, VMPs were enriched for model for the same CpG only and did not require that the
distinct genomic features, including intergenic regions same SNPs be included in the model across cohorts. The
(p < 2.20 × 10−16, OR = 1.65, Fisher’s test) and enhancers majority of VMPs (43.87%) were consistently best
(p < 2. 20 × 10−16, OR = 1.87, Fisher’s test). explained by G × CA models. These results remained
We examined whether interindividual differences in stable with inclusion of cohort-specific covariates,
DNAm levels of overlapping VMPs were better explained including symptoms severity (see Supplementary Figs. 4
by genotype in cis (defined as 1 MB window around the and 5). For all cohorts, ΔadjR2, i.e., the difference between
specific VMP), by CA (E), or by additive or interaction the adjusted R2 of the best models to the adjusted R2 ofCzamara et al. Translational Psychiatry (2021)11:88 Page 6 of 11
Fig. 2 Enrichment of overlapping VMPs with regard to ChromHMM states. Histone mark enrichment for overlapping VMPs with regard to other
450K CpG sites (above panel) (a). Histone mark enrichment for overlapping VMPs (below) with at least two consistent best G, G + CA, or G × CA
models against overlapping VMPs with no consistent models. Green color indicates depletion, and red color indicates enrichment. Thick black lines
around the rectangles indicate significant enrichment/depletion based on Fisher’s tests and a Bonferroni threshold of p < 4.66 × 10−04. VMPs with at
least two consistent best G × CA models were enriched in repressed Polycomb (p = 5.48 × 10−19, OR = 1.19, Fisher’s test) (b). Average distance
between SNP and CpG site for consistent best G (left), consistent best G + CA (middle), and consistent G × CA models (right). CpGs with at least two
consistent G × CA models presented with a significantly longer distance between SNP and VMP than VMPs with other consistent models (mean
absolute distance = 271.83 kb, p < 2.20 × 10−16, Wilcoxon’s test) (c).
the next best model, was highest for VMPs where G × CA HBCS, a cohort of 151 elderly individuals, of which 77 had
was chosen as best model and significantly larger as been evacuated to Sweden or Denmark during World
compared to G and G + CA models (see Supplementary War II.
Fig. 6A–C, p < 2.2 × 10−16 for all cohorts, Wilcoxon’s To base the comparison of best models across the
test). VMPs with at least two consistent best G × CA developmental trajectory on the same CpG sites in all
models were enriched in repressed Polycomb (see Fig. 2a, cohorts, we used the overlap of VMPs identified in
b) and presented with a significantly longer distance GRADY/PReDICT/U19 and CpGs available in BerlinLCS
between SNP and VMP than VMPs with other consistent as well as in HBCS. This resulted in 36,091 VMPs avail-
models (see Fig. 2c). able in all five cohorts. Even with this more restricted set
of VMPs, the best models remained combined models of
VMPs across the life course and with documented G and CA (see Fig. 3a and Supplementary Fig. 7).
adversity There were 1672 VMPs (5.4%) with a consistent best
To test if the identified combined effects of genotype model across all five cohorts (see Fig. 3b). Among these
and CA are stable across the life course and also observed stable VMPs, 83.73% had G × CA as the best model (n =
with documented and not only self-reported adversity, we 1400, “stable across age G × CA CpGs”).
used two additional cohorts. The BerlinLCS cohort, In comparison, only 670 VMPs were consistently best
consisting of 173 DNAm saliva samples of children aged explained by G × CA across the three adult cohorts but
between 3 and 5 years, of which 86 were recruited from neither in the BerlinLCS nor in the HBCS (“stable in
social services and other child welfare centers due to MT adults G × CA CpGs”). Both groups of CpGs were sig-
or neglect. At the other end of the age spectrum is the nificantly enriched (p < 0.002 for both groups of CpGs) forCzamara et al. Translational Psychiatry (2021)11:88 Page 7 of 11
Fig. 3 Distribution of the best models explaining variation in DNAm across the five cohorts. Percentage of overlapping VMPs (n = 36,091) best
explained by G, CA, G + CA, or G × CA in each cohort using the highest adjusted R2 For HBCS, 49.03% of VMPs presented with best model G × CA,
37.40% with best model G, and 13.55% with best model G + CA, and for BerlinLCS, 61.26% of VMPs presented with best model G × CA, 29.51% with
best model G, and 8.75% with G + CA (a). Consistency of best model explaining DNAm variability in VMPs across five cohorts. Percentage of VMPs
with the same model explaining variation in DNAm best overlapping between the cohorts stratified by model type (b).
eQTM sites49 as compared to all 450K CpGs (based on Stable across age G × CA CpGs were significantly enri-
10,000 randomly drawn CpG sets). ched for 24 GO terms, and stable in adults G × CA CpGs
were significantly enriched for 35 GO terms in the bio-
Tissue specificity and functions of genes linked to stable logical processes categories (all FDR-corrected p values <
G × CA CpGs 0.05). While some of these processes overlapped, stable
We annotated each VMP to the closest gene and used across age G × CA CpGs were selectively significantly
the list of unique genes to test for differences in gene-set enriched in categories reflecting processes related to
enrichment using FUMA45. As a background set, we neuron development and synapse organization (see Sup-
mapped the 36,091 VMPs which were used in the analysis plementary Fig. 8). Stable across age G × CA CpGs were
across all five cohorts representing 10,308 unique genes. significantly enriched for the cellular component terms
We tested gene lists derived from stable across age G × “neuron part” and “neuron projection” and the molecular
CA CpGs and stable in adults G × CA CpGs for enrich- function terms “DNA binding transcription factor activ-
ment in differentially expressed gene sets across different ity,” “sequence specific DNA binding,” and “sequence
tissues using the GTEx database46. The genes mapping to specific double DNA binding” (all FDR-corrected p values
stable across age G × CA CpGs (1400 CpGs mapping to < 0.05). Stable in adults G × CA CpGs had no cellular
1123 unique genes) were significantly enriched for genes component or molecular function term significantly
specific to brain (FDR-corrected p value = 2.85 × 10−04) enriched. Non-consistent CpGs showed no significant
but not to other tissues. This enrichment for brain tran- consistent enrichments for any GO terms.
scripts was not observed for genes mapped to stable in The analyses investigating tissue-specific gene expres-
adults G × CA CpGs (670 CpGs mapping to 584 unique sion as well as GO terms point to the fact that stable G ×
genes, FDR-corrected p value = 1.00 × 10−01). As control, CA VMPs could have a distinct functional relevance,
we compared these to enrichments from random subsets related to development and brain function.
from the list of inconsistent CpGs that showed different
best models across all three adult groups (see Fig. 1c, Discussion
n = 5652). We randomly picked groups of 1123 genes In this study, we investigated the contributions of
(which is the number of genes matching to stable across exposure to CA, genotype in cis as well as their additive
age G × CA CpGs) matching to these CpGs. None of these and interactive effects on interindividual variability of
subsets showed significant tissue-specific enrichments DNAm in variable CpGs in peripheral tissues. Inde-
(see Fig. 4). pendent of the age of the cohort, we observed thatCzamara et al. Translational Psychiatry (2021)11:88 Page 8 of 11 Fig. 4 Enrichment of stable across age G × CA, stable in adults G × CA and non-consistent CpGs for GTEx upregulated gene sets. The x-axis denotes the −log10(p value), and the y-axis the specific tissue type. Significant enrichments (based on FDR correction of 0.05) are depicted in blue, and nonsignificant in gray. For the group of nonconsistent CpGs, we randomly picked ten subsets of genes mapping to nonconsistent CpGs, equal in number to the genes mapping to stable across age G × CA CpGs. None of these subsets presented with significant adjusted p values for enrichment. In the plot, median p values across all subsets are depicted. models combining G and CA best explained DNAm results for the effect of CA alone50,51. Indeed, very few of variability in the majority of CpGs, suggesting that the the overlapping VMPs were best explained by environ- extent of the combined impact on DNAm is similar for ment independent of genotype (
Czamara et al. Translational Psychiatry (2021)11:88 Page 9 of 11
The proportions of best G × CA models which we sufficient power to detect consistent effect directions after
identified in our cohorts are analogous to the ~70% of correction for multiple testing. In order to have sufficient
variably methylated regions that were shown to be best power to assess specific SNP × CA effects surviving mul-
explained by integrated genetic and prenatal environment tiple testing correction larger, ethnically homogenous
effects in neonates16,17. Our findings corroborate that cohorts are necessary. All reported interactions are sta-
genotype acts as an important moderating influence on tistical interactions and limited to a cis window around
main environmental effects on DNAm also in the context the CpG site. Further experiments are required to assess
of a postnatal stressor. In fact, G × CA was the model that whether these would also reflect biological/mechanistic
best explained DNAm variance in the majority (83%) of interactions. Along the same lines, much larger cohorts
CpGs that had the same best model across cohorts. The will be needed to assess potential trans effects.
stability of the best model for these CpGs cuts across a Furthermore, strategies to reduce the number of tests,
large age range from early childhood to late adulthood, i.e., SNPs, are needed. Possible methods include the pre-
across tissue (blood and saliva), different DNAm varia- filtering for functionally relevant SNPs using deep learn-
bility thresholds, psychiatric diagnoses, as well as self- ing algorithm such as DeepSEA for instance54, or
reported retrospective vs. verified CA. Additional studies experimental approaches such as SNPs disrupting tran-
in longitudinal cohorts with repeated measures of DNAm scription factor binding or chromatin structure55. How-
in the same individuals are needed to confirm such sta- ever, our results can highlight those CpGs that are most
bility across time. influenced by the combination of a genetic variant in cis
The VMPs with stable G × CA models mapped to genes and CA in a consistent manner across five cohorts and
with distinct functionality. In contrast to VMPs that only hence are environmentally sensitive. While our results
showed the G × CA model in all adult cohorts, but not highlight convergent effects of CA across ages, we did not
more, the genes mapped to stable G × CA VMPs across have sufficient power to identify effects specific to certain
five cohorts were enriched for transcripts specific to the forms of CA or neglect, or related to specific timing of the
brain (see Fig. 4) as well as to GO terms related to brain exposure. Our analysis provides a possible framework of
development and synapse function. Importantly, G × CA how specific combined effects of genotype and environ-
VMPs were also enriched for eQTMs, indicating that any ment on DNAm might be studied in the future.
factors influencing variability at these loci will have effects In conclusion, in this study, we show that CA has a
on gene transcription. Our samples size was under- larger impact on DNAm in combination with genetic
powered to reliably detect consistent effect directions of variation than by itself. Inclusion of information on
SNP × CA interactions after correction for multiple test- genetic variation may thus help to uncover impact of
ing. For this larger, ethnically homogenous cohorts will be environmental factors on epigenetic measures that would
necessary. Nonetheless, our results can highlight those otherwise remain concealed. Such combined approaches
CpGs that are most influenced by the combination of a could support to identify gene pathways relevant to risk or
genetic variant in cis and CA in a consistent manner resilience following exposure to CA.
across age, unmasking an epigenetic signature of CA.
Consistent with the previous literature14, cis meQTLs Acknowledgements
The authors would like to thank all study participants as well as all involved in
were clearly apparent in the five independent cohorts with the HBCS. The authors also acknowledge the Genetics Core of the Wellcome
diverse ethnic backgrounds. Although it is known that Trust Clinical Research Facility (Edinburgh, UK) which used the 450K array for
Caucasian and African American meQTLs significantly the HBCS samples. The PReDICT study was supported by the following
National Institutes of Health grants: P50 MH077083; R01 MH080880; UL1
overlap, 14–45% shows specificity for ethnicity52. In our RR025008; and M01 RR0039. Funding for the U19 study was provided from a
analysis, we found converging evidence that G × CA grant from the National Institute of Mental Health, U19 MH069056, with
interactions best explain variability of DNAm across dif- additional support from VA CSRD Project ID 09S-NIMH-002. The GRADY study
was supported by a grant from the National Institute of Mental Health, R01
ferent ethnicities, but this does not exclude ancestry MH071537-01A1. HBCS has been supported by grants from the British Heart
specific effects. To identify such specific interaction Foundation, Academy of Finland, the Finnish Diabetes Research Society,
effects, larger samples for each ethnicity are required. Folkhälsan Research Foundation, Novo Nordisk Foundation, Finska
Läkaresällskapet, Signe and Ane Gyllenberg Foundation, University of Helsinki,
Finally, we want to note the limitations of this study. Ministry of Education, Ahokas Foundation, and the Emil Aaltonen Foundation.
First, we restricted our analyses to specific DNAm array The BerlinLCS study was funded by a research grant from the German Federal
contents and to potentially functional CpGs, i.e., VMPs, Ministry of Education and Research (BMBF) to CMH and EBB (FKZ 01KR1301B).
AJD has received a Scottish Senior Clinical Fellowship (SCD/09).
so that we do not reflect every CpG tested on the array.
Second, we used the adjusted R2 as main criterion for Author details
model fit as we were mainly interested in explaining 1
Department of Translational Research in Psychiatry, Max Planck Institute of
variability of DNAm. A variety of other model selection Psychiatry, Kraepelinstr. 2-10, 80804 Munich, Germany. 2Department of
Complex Trait Genetics, Center for Neurogenomics and Cognitive Research,
criteria are available53 and which one to choose is an Amsterdam Neuroscience, Vrije Universiteit Amsterdam, 1081 HV Amsterdam,
ongoing debate. Third, our analysis does not provide The Netherlands. 3Department of Psychology and Logopedics, Faculty ofCzamara et al. Translational Psychiatry (2021)11:88 Page 10 of 11
Medicine, University of Helsinki, Haartmaninkatu 3, 00014 Helsinki, Finland. consultant for Aptinyx, Greenwich Biosciences, Myriad Neuroscience, Otsuka,
4
Sopra Steria, 80804 Munich, Germany. 5University/British Heart Foundation and Sophren Therapeutics. HSM receives consulting and intellectual property
Centre for Cardiovascular Science, Queen’s Medical Research Institute, licensing fees from Abbott Neuromodulation. SJM is supported through the
Edinburgh BioQuarter, University of Edinburgh, 47 Little France Crescent, use of facilities and resources at the Michael E. Debakey VA Medical Center,
Edinburgh EH16 4TJ, UK. 6Institute for Molecular Medicine Finland (FIMM), Houston, TX, and has served as a consultant to Alkermes, Allergan, Signant
University of Helsinki, 00014 Helsinki, Finland. 7Department of Child and Health, Clexio Biosciences, Janssen, Neurocrine, Perception Neurosciences,
Adolescent Psychiatry, Charité—Universitätsmedizin Berlin, Campus Virchow, Praxis Precision Medicines, Sage Therapeutics, and Seelos Therapeutics. He has
13353 Berlin, Germany. 8Department of Psychiatry, Dell Medical School, received research support from Biohaven Pharmaceuticals and VistaGen
University of Texas at Austin, 1601 Trinity St, Austin, TX 78712, USA. Therapeutics. In the past 5 years, DVI has received consulting fees from
9
Department of Psychiatry and Behavioral Sciences, Emory University School of Alkermes, Axsome, Centers for Psychiatric Excellence, Global Medical
Medicine, 12 Executive Park Dr, Atlanta, GA 30329, USA. 10Icahn School of Education, MYnd Analytics (CNS Response), Jazz, Lundbeck, Otsuka, Precision
Medicine at Mount Sinai, 1 Gustave L. Levy PI, New York, NY 10029, USA. Neuroscience, Sage, and Sunovion, and has received research support
11
Menninger Department of Psychiatry and Behavioral Sciences, Baylor College (through his academic institutions) from Alkermes, Astra Zeneca, BrainsWay,
of Medicine Mental Health Care Line, Michael E. Debakey VA Medical Center, LiteCure, NeoSync, Roche, Shire, and Otsuka. TCN has received study
Houston, TX, USA. 12NYU School of Medicine and Nathan Kline Institute, New medication from Corcept Therapeutics and served as a consultant for Jazz
York, NY, USA. 13Departments of Psychiatry and Neurology, University of Pharmaceuticals. Over the last 5 years, DVI has received consulting fees from
California, San Francisco, CA, USA. 14Charité–Universitätsmedizin Berlin, Axsome, Alkermes, Centers of Psychiatric Excellence, MYnd Analytics (CNS
Corporate Member of Freie Universität Berlin, Humboldt-Universität zu Berlin, Response), Jazz, Lundbeck, Precision Neuroscience, Otsuka, and Sunovion, and
Berlin Institute of Health (BIH), Institute of Medical Psychology, Luisenstraße 57, has received research support (through his academic institutions) from
10117 Berlin, Germany. 15Turku Institute for Advanced Studies, University of Alkermes, Astra Zeneca, BrainsWay, LiteCure, NeoSync, Roche, and Shire. EBB is
Turku, 20500 Turku, Finland. 16Department of General Practice and Primary the coinventor of FKBP5: a novel target for antidepressant therapy, European
Health Care, Helsinki University Hospital, University of Helsinki, 00290 Helsinki, Patent no. EP 1687443 B1, and receives a research grant from Böhringer
Finland. 17Folkhälsan Research Center, 00250 Helsinki, Finland. 18Department of Ingelheim for a collaboration on functional investigations of FKBP5.
Obstetrics and Gynecology, Yong Loo Lin School of Medicine, National
University of Singapore, Singapore, Singapore. 19Singapore Institute for Clinical
Sciences, Singapore, Singapore. 20Mailman Research Center, 115 Mill St., Publisher’s note
Mailstop 339, Belmont, MA 02478, USA. 21Faculty of Health Sciences, Simon Springer Nature remains neutral with regard to jurisdictional claims in
Fraser University, 8888 University Drive, Burnaby, BC, Canada. 22BC Children’s published maps and institutional affiliations.
Hospital Research Institute, Vancouver, BC, Canada
Supplementary Information accompanies this paper at (https://doi.org/
Funding 10.1038/s41398-020-01147-z).
Open Access funding enabled and organized by Projekt DEAL.
Received: 19 November 2020 Revised: 2 December 2020 Accepted: 4
Data availability December 2020
Due to ethical issues and consent, datasets from PReDICT, U19, BerlinLCS, and
HBCS analyzed during the current study are not publicly available. Interested
researchers can obtain a deidentified dataset after approval from the
respective study boards. Data requests may be subject to further review by the
national register authority and the ethical committees (HBCS). The raw
methylation data and all related phenotypes for the GRADY cohort have been References
deposited into NCBI GEO (GSE72680). 1. Kessler, R. C. et al. Childhood adversities and adult psychopathology in
the WHO World Mental Health Surveys. Br. J. Psychiatry 197, 378–385
Conflict of interest (2010).
DC, ET, JT, JM, YA, AJD, BK, AP, SMW, BK, DVI, CMH, JL, JGE, KR, KJR, and NP 2. Kosidou, K. et al. Socioeconomic status and risk of psychological distress and
report no financial disclosures. CBN: Research/Grants: National Institutes of depression in the Stockholm Public Health Cohort: a population-based study.
Health; consulting (last 3 years): Xhale, Takeda, Taisho Pharmaceutical Inc., J. Affect. Disord. 134, 160–167 (2011).
Signant Health, Sunovion Pharmaceuticals Inc., Janssen Research & 3. Luby, J. L., Barch, D., Whalen, D., Tillman, R. & Belden, A. Association between
Development LLC, Magstim, Inc., Navitor Pharmaceuticals, Inc., Sunovion, TC early life adversity and risk for poor emotional and physical health in ado-
MSO, Inc., Intra-Cellular Therapies, Inc., EMA Wellness, Gerson Lehrman Group, lescence: a putative mechanistic neurodevelopmental pathway. JAMA Pediatr.
and Acadia Pharmaceuticals; stockholder: Xhale, Celgene, Seattle Genetics, 171, 1168–1175 (2017).
AbbVie, OPKO Health, Inc., Antares, BI Gen Holdings, Inc., Corcept Therapeutics, 4. Molnar, B. E., Buka, S. L. & Kessler, R. C. Child sexual abuse and subsequent
TC MSO, Inc., Trends in Pharma Development, LLC, and EMA Wellness; scientific psychopathology: results from the National Comorbidity Survey. Am. J. Public
advisory boards: American Foundation for Suicide Prevention (AFSP), Brain and Health 91, 753–760 (2001).
Behavior Research Foundation, Xhale, Anxiety Disorders Association of America 5. Teicher, M. H. & Khan, A. Childhood maltreatment, cortical and amygdala
(ADAA), Skyland Trail, Signant Health, and Laureate Institute for Brain Research morphometry, functional connectivity, laterality, and psychopathology. Child
(LIBR), Inc.; board of directors: AFSP, Gratitude America, ADAA, and Xhale Smart, Maltreatment 24, 458–465 (2019).
Inc.; income sources or equity of $10,000 or more: American Psychiatric 6. Pesonen, A. K. et al. Childhood separation experience predicts HPA axis hor-
Publishing, Xhale, Signant Health, CME Outfitters, Intra-Cellular Therapies, Inc., monal responses in late adulthood: a natural experiment of World War II.
Magstim, and EMA Wellness; patents: Method and devices for transdermal Psychoneuroendocrinology 35, 758–767 (2010).
delivery of lithium (US 6,375,990B1), Method of assessing antidepressant drug 7. Segerstrom, S. C. & Miller, G. E. Psychological stress and the human immune
therapy via transport inhibition of monoamine neurotransmitters by ex vivo system: a meta-analytic study of 30 years of inquiry. Psychol. Bull. 130, 601–630
assay (US 7,148,027B2), and Compounds, Compositions, Methods of Synthesis, (2004).
and Methods of Treatment (CRF Receptor Binding Ligand) (US 8,551, 996 B2). 8. Szyf, M. & Bick, J. DNA methylation: a mechanism for embedding early life
WEC is a board member of Hugarheill ehf, an Icelandic company dedicated to experiences in the genome. Child Dev. 84, 49–57 (2013).
the prevention of depression, and he receives book royalties from John Wiley 9. Gutierrez-Arcelus, M. et al. Passive and active DNA methylation and the
& Sons. His research is also supported by the Mary and John Brock Foundation interplay with genetic variation in gene regulation. Elife 2, e00523 (2013).
and the Fuqua family foundations. He is a consultant to the George West 10. Cecil, C. A. M., Zhang, Y. & Nolte, T. Childhood maltreatment and DNA
Mental Health Foundation and is a member of the Scientific Advisory Board of methylation: a systematic review. Neurosci. Biobehav. Rev. 112, 392–409
the ADAA and the AIM for Mental Health Foundation. BWD has received (2020).
research support from Acadia, Aptinyx, Axsome, Compass Pathways, 11. Ziller, M. J. et al. Charting a dynamic DNA methylation landscape of the
Intracellular Therapies, Janssen, Sage, and Takeda. He has served as a human genome. Nature 500, 477–481 (2013).Czamara et al. Translational Psychiatry (2021)11:88 Page 11 of 11
12. Gu, J. et al. Mapping of variable DNA methylation across multiple cell types 33. Houseman, E. A. et al. DNA methylation arrays as surrogate measures of cell
defines a dynamic regulatory landscape of the human genome. G3 6, mixture distribution. BMC Bioinform. 13, 86 (2012).
973–986 (2016). 34. Smith, A. K. et al. DNA extracted from saliva for methylation studies of psy-
13. Portela, A. & Esteller, M. Epigenetic modifications and human disease. Nat. chiatric traits: evidence tissue specificity and relatedness to brain. Am. J. Med.
Biotechnol. 28, 1057–1068 (2010). Genet. B Neuropsychiatr. Genet. 168B, 36–44 (2015).
14. Gibbs, J. R. et al. Abundant quantitative trait loci exist for DNA methylation and 35. Elliott, H. R. et al. Differences in smoking associated DNA methylation patterns
gene expression in human brain. PLoS Genet. 6, e1000952 (2010). in South Asians and Europeans. Clin. Epigenetics 6, 4 (2014).
15. Gaunt, T. R. et al. Systematic identification of genetic influences on methyla- 36. Zeilinger, S. et al. Tobacco smoking leads to extensive genome-wide
tion across the human life course. Genome Biol. 17, 61 (2016). changes in DNA methylation. PLoS ONE 8, https://doi.org/10.1371/journal.
16. Teh, A. L. et al. The effect of genotype and in utero environment on inter- pone.0063812 (2013).
individual variation in neonate DNA methylomes. Genome Res. 24, 1064–1074 37. Richmond, R. C., Suderman, M., Langdon, R., Relton, C. L. & Davey Smith, G.
(2014). DNA methylation as a marker for prenatal smoke exposure in adults. Int J.
17. Czamara, D. et al. Integrated analysis of environmental and genetic influences Epidemiol. 47, 1120–1130 (2018).
on cord blood DNA methylation in new-borns. Nat. Commun. 10, 2548 (2019). 38. Purcell, S. et al. PLINK: a tool set for whole-genome association and
18. Davis, R. G., Ressler, K. J., Schwartz, A. C., Stephens, K. J. & Bradley, R. G. population-based linkage analyses. Am. J. Hum. Genet. 81, 559–575 (2007).
Treatment barriers for low-income, urban African Americans with undiag- 39. Howie, B. N., Donnelly, P. & Marchini, J. A flexible and accurate genotype
nosed posttraumatic stress disorder. J. Trauma Stress 21, 218–222 (2008). imputation method for the next generation of genome-wide association
19. Gillespie, C. F. et al. Trauma exposure and stress-related disorders in inner city studies. PLoS Genet. 5, e1000529 (2009).
primary care patients. Gen. Hosp. Psychiatry 31, 505–514 (2009). 40. Genomes Project, C. et al. A map of human genome variation from
20. Zannas, A. S. et al. Lifetime stress accelerates epigenetic aging in an urban, population-scale sequencing. Nature 467, 1061–1073 (2010).
African American cohort: relevance of glucocorticoid signaling. Genome Biol. 41. Bernstein, D. P. et al. Initial reliability and validity of a new retrospective
16, 266 (2015). measure of child abuse and neglect. Am. J. Psychiatry 151, 1132–1136 (1994).
21. Zannas, A. S. et al. Correction to: lifetime stress accelerates epigenetic aging in 42. Barnett, D., Manly, J. T. & Cicchetti, D. In Child abuse, child development, and
an urban, African American cohort: relevance of glucocorticoid signaling. social policy (Ciccetti, D. & Toth, S. L. eds) (Ablex, 1993).
Genome Biol. 19, 61 (2018). 43. Egger, H. L. & Angold, A. In Handbook of Infant, Toddler, and Preschool Mental
22. Dunlop, B. W. et al. Predictors of remission in depression to individual and Health Assessment (eds DelCarmen-Wiggins, R. & Carter, A. S.) 223–243 (Oxford
combined treatments (PReDICT): study protocol for a randomized controlled University Press, 2004).
trial. Trials 13, 106 (2012). 44. Yu, G., Wang, L. G. & He, Q. Y. ChIPseeker: an R/Bioconductor package for ChIP
23. Dunlop, B. W. et al. Corticotropin-releasing factor receptor 1 antagonism is peak annotation, comparison and visualization. Bioinformatics 31, 2382–2383
ineffective for women with posttraumatic stress disorder. Biol. Psychiatry 82, (2015).
866–874 (2017). 45. Jaffe, A. E. et al. Bump hunting to identify differentially methylated regions in
24. Barker, D. J., Osmond, C., Forsen, T. J., Kajantie, E. & Eriksson, J. G. Trajectories of epigenetic epidemiology studies. Int J. Epidemiol. 41, 200–209 (2012).
growth among children who have coronary events as adults. N. Engl. J. Med. 46. Watanabe, K., Taskesen, E., van Bochoven, A. & Posthuma, D. Functional
353, 1802–1809 (2005). mapping and annotation of genetic associations with FUMA. Nat. Commun. 8,
25. Eriksson, J. G., Osmond, C., Kajantie, E., Forsen, T. J. & Barker, D. J. Patterns of 1826 (2017).
growth among children who later develop type 2 diabetes or its risk factors. 47. Consortium, G. T. The Genotype-Tissue Expression (GTEx) project. Nat. Genet.
Diabetologia 49, 2853–2858 (2006). 45, 580–585 (2013).
26. Räikkönen, K. et al. Infant growth and hostility in adult life. Psychosom. Med. 70, 48. Theil, H. Economic Forecasts and Policy, 213 (North-Holland Pub. Co., Amster-
306–313 (2008). dam, 1961).
27. Khulan, B. et al. Epigenomic profiling of men exposed to early-life stress 49. Bonder, M. J. et al. Disease variants alter transcription factor levels and
reveals DNA methylation differences in association with current mental state. methylation of their binding sites. Nat. Genet. 49, 131–138 (2017).
Transl. Psychiatry 4, e448 (2014). 50. Marzi, S. J. et al. Analysis of DNA methylation in young people: limited evi-
28. Aryee, M. J. et al. Minfi: a flexible and comprehensive bioconductor package dence for an association between victimization stress and epigenetic variation
for the analysis of infinium DNA methylation microarrays. Bioinformatics 30, in blood. Am. J. Psychiatry 175, 517–529 (2018).
1363–1369 (2014). 51. Lang, J. et al. Adverse childhood experiences, epigenetics and telomere length
29. Fortin, J. P. et al. Functional normalization of 450k methylation array data variation in childhood and beyond: a systematic review of the literature. Eur.
improves replication in large cancer studies. Genome Biol. 15, 503 (2014). Child Adolesc. Psychiatry, https://doi.org/10.1007/s00787-019-01329-1 (2019).
30. Johnson, W. E., Li, C. & Rabinovic, A. Adjusting batch effects in microarray 52. Smith, A. K. et al. Methylation quantitative trait loci (meQTLs) are consistently
expression data using empirical Bayes methods. Biostatistics 8, 118–127 detected across ancestry, developmental stage, and tissue type. BMC Genom.
(2007). 15, 145 (2014).
31. Leek, J. T., Johnson, W. E., Parker, H. S., Jaffe, A. E. & Storey, J. D. The sva package 53. Zhang, P. Inference after variable selection in linear regression models. Bio-
for removing batch effects and other unwanted variation in high-throughput metrika 79, 741–746 (1992).
experiments. Bioinformatics 28, 882–883 (2012). 54. Zhou, J. & Troyanskaya, O. G. Predicting effects of noncoding variants with
32. Chen, Y. A. et al. Discovery of cross-reactive probes and polymorphic CpGs in deep learning-based sequence model. Nat. Methods 12, 931–934 (2015).
the Illumina Infinium HumanMethylation450 microarray. Epigenetics 8, 55. Schoenfelder, S. & Fraser, P. Long-range enhancer-promoter contacts in gene
203–209 (2013). expression control. Nat. Rev. Genet. 20, 437–455 (2019).You can also read
